Integrating molecular representations for machine learning-based virtual screening of Glutaminyl Cyclase inhibitors

Secretory Glutaminyl Cyclase (sQC) catalyzes the formation of pyroglutamate amyloid-β (pE-Aβ), a highly aggregation-prone and neurotoxic species. In this study, we developed a deep learning model that combines ChemBERTa pre-trained embeddings with extended-connectivity fingerprints (ECFP) to predict IC₅₀ values of compounds targeting sQC, using data from the ChEMBL database. The model's performance was systematically compared with several traditional machine learning algorithms trained on different molecular representations, including ECFP fingerprints, embeddings from ChemBERTa and MolFormer, 2D and 3D molecular descriptors, and their combinations. The best-performing model was subsequently applied to screen natural products from the COCONUT database. Potential compounds were further evaluated through ADME analysis and molecular docking to identify those with favorable pharmacokinetic properties and high affinity for sQC. Three representative candidates (CNP0534898.2, CNP0421664.1, and CNP0273039.1) were selected for molecular dynamics simulations. The stability of protein-ligand interactions observed during the simulations, together with binding free energy estimation using MM/GBSA method, further supports the reliability of the proposed workflow for identifying potential natural product-derived sQC inhibitors.